| 面向地学应用的遥感影像分割评价 |
| |
| 引用本文： | 张仙,明冬萍.面向地学应用的遥感影像分割评价[J].测绘学报,2015(Z1):108-116. |
| |
| 作者姓名： | 张仙 明冬萍 |
| |
| 作者单位： | 中国地质大学 北京 信息工程学院,北京,100083 |
| |
| 基金项目： | 国家自然科学基金(41371347)，中央高校基本科研业务费专项资金Foundation support:National NaturalScience Foundation of China(41371347)，The Fundamental Research Funds for the Central Universities |
| |
| 摘 要： | 影像分割是面向地理对象影像分析(GEOBIA)中的一个关键环节。分割评价有助于为影像选择合适的分割方法和最佳分割尺度。本文提出一种以遥感地学应用面向的对象为依据的分割方法分类及评价体系。首先将影像分割方法分为面向局部特征监测的典型目标识别和面向全局特征监测的面向GEOBIA的分割方法两组,进而针对这两组分割方法提出了两套分割评价测度指标及相应的综合评价方法。在面向典型目标识别的分割方法评价中,使用区域内部非均质度、区域间灰度对比度、区域间散度对比度、边界点梯度和单位像素运行时间作为评价测度,并针对由于评价测度间的相关性而无法直接确定各测度权重分配的问题,提出利用熵权法为各个评价测度分配权重以获得综合评价结果的分割评价方法,该评价方法可用于选择合适的分割方法。用于面向GOEBIA的分割方法中使用分割区域内均质性和区域间异质性作为评价测度,这种评价方法适用于选择最优分割尺度参数。本文通过定量试验论证了这两种评价方法的有效性,试验结果表明其在遥感应用中具有实际意义。最后本文分析了影像分割评价方法的不足及未来的发展方向。
|
| 关 键 词： | 影像分割评价 目标识别 评价测度 |
Geo-appl ication Oriented Evaluations of Remote Sensing Image Segmentation |
| |
| Abstract: | Image segmentation is a key technique in information extraction from remote sensing images and it is especially the foundational procedure of geographic object-based image analysis (GEOBIA)in which the digital image is transformed from discrete pixels into homogeneous image object primitives. However,facing with large amount of image segmentation methods and various ultimate objects of remote sensing application,evaluation of image segmentation is very significant and it is helpful for selecting suitable segmentation method and optimal segmentation parameters.Other than traditional classification of image segmentation methods based on the principle of algorithms but based on ultimate objects of remote sensing application,this paper firstly divides the image segmentation methods into two groups.One group is for typical object recognition,which is designed for local feature detection.Another group is for object oriented image classification,which is designed for global features detections.Then this paper proposes two sets of segmentation evaluation measurement indexes and correspondingly introduces their principles and computational formulae for both two groups.The measurement indexes used in segmentation for typical object recognition are intra-region non-uniformity,inter-region gray-level contrast,inter-region variance contrast,edge gradient and running time per-pixel.However,the weight allocation can not be determined straightly because of the correlation of measures.To resolve this problem,this paper proposes a segmentation evaluation method that distribute weight for measures with entropy method to evaluate the segmentation method comprehensively.This method is useful for selecting suitable segmentation method. The measurement indexes used in segmentation for object based image classification are homogeneity and heterogeneity of the segmentation image.This method is useful for selecting optimal segmentation parameters.This paper demonstrates and verifies the two kinds of evaluations by quantitative experiments. The quantitative experimental results show that the two image segmentation evaluation methods are practically meaningful in remote sensing applications.In the end,this paper analyzes the defects and feature improvements of image segmentation evaluation methods. |
| |
| Keywords: | image segmentation evaluation object recognition measurement index |
| 本文献已被 CNKI 万方数据 等数据库收录！ |
|
